Comment by cprayingmantis

I noticed something similar with images as inputs to Claude, you can scale down the images and still get good outputs. There is an accuracy drop off at a certain point but the token savings are worth doing a little tuning there.